Within one year: 6 new payment protocols + 3 HTTP 402 revival initiatives + IETF standardization—all driven by a single common catalyst: AI agents beginning to have their own payment needs.

Article author and source: Yajin's Scientific Research and Entrepreneurship Journal

On April 29, 2025, Mastercard held a launch event at its headquarters in Purchase, New York, officially announcing the "Agent Pay" product [3]. There was little fanfare that day, and most in the industry treated it as just another routine announcement from a card network.

Over the next 12 months, things became unusual.

In May, Coinbase launched x402, finally activating the HTTP status code 402 that had been reserved since 1998 but never used, enabling USDC transfers to be directly embedded in HTTP responses [1]. Two months later, Cloudflare deployed Pay per Crawl using the same status code, introducing pay-per-use pricing for AI crawlers [9], signaling that the idea of “using HTTP 402 for payments” had spread from the crypto community to content providers. Two months after that, protocol layers entered the fray: in September, Google partnered with over 60 companies to launch AP2, while OpenAI and Stripe simultaneously introduced ACP, enabling ChatGPT to complete purchases on Etsy on behalf of users directly within a conversation [2][4]. In October, card networks joined in: Visa and Cloudflare jointly released TAP, granting agents their first verifiable card network identity [5].

Entering 2026, the entire industry transitioned from a fragmented emergence to a standardized phase. In February, Lightning Labs applied the L402 concept to Bitcoin, creating L402 [8]. In March, Stripe and Paradigm jointly released MPP, while Tempo launched on mainnet; just 12 days later, the full solution was submitted to the IETF standardization process [6]. In April, the Linux Foundation took over the x402 Foundation, with over 20 founding members—including card networks (Visa, Mastercard, American Express), cloud providers (AWS, Google, Microsoft), and crypto infrastructure firms (Coinbase, Circle, Stripe)—all seated together, representing three previously competing camps [1]. This marked the first time a payment protocol received simultaneous endorsement from all three of these groups.

Within one year: 6 new payment protocols + 3 HTTP 402 revival initiatives + IETF standardization—all driven by a single common catalyst: AI agents beginning to have their own payment needs.

The entire赛道 spawned by this new requirement is now uniformly referred to in the industry as Agentic Payment: enabling agents to act as independent payment entities that are identified, authorized, settled, and held accountable. The fundamental assumption that “the final decision rests with a human”—which has held for 60 years—has for the first time begun to change.

This five-part series breaks down this very issue. In this installment, we first clarify the context: where the traditional payment stack is stuck, and which gaps each of the three forces is targeting. The next three articles will each dive deep into one of the three engineering pathways, and the final article will provide a cross-comparison.

Why traditional payment stacks are unsuitable for agent payments

To understand why protocols are experiencing a surge in concentration, let’s first examine how money flows in a traditional consumer payment.

For example, Xiao Shuai, a running enthusiast, uses his Visa card to buy running shoes on Nike’s website. The card number is issued by an issuer (possibly Chase). Nike, as the merchant, cannot process card payments directly and must route the transaction to an acquirer (such as Adyen). Adyen sends the transaction through the Visa network to Chase. Chase deducts $100 from Xiao Shuai’s account; part of this amount is retained as fees: for a $100 transaction, the issuer takes approximately 1.5% (called interchange), Visa collects about 0.1% (called network fee), and Adyen takes around 1.5% (called processing fee). In the end, approximately $96 reaches Nike’s account.

The 3-4% is not paid by Xiao Shuai—it’s paid by Nike (Xiao Shuai’s card is still charged $100, but Nike actually receives $96). Why is Nike willing to bear this cost? Because over the past sixty years, card networks have accumulated an almost irreplaceable asset: chargebacks. This system provides a safety net for consumers like Xiao Shuai, giving them the confidence to shop online and from unfamiliar merchants. If Xiao Shuai discovers he received the wrong shoe model, Nike fails to ship, or an unauthorized charge appears on his card, he can call Chase to initiate a dispute. A decades-old adjudication process involving the issuing bank, Visa, Adyen, and Nike then resolves the issue—and if successful, Xiao Shuai gets his money back. Nike pays the 3-4% to buy this trust.

The same applies to another market. When Chinese consumers use a China Merchants Bank card or scan Alipay to buy running shoes on JD.com, the rail structure looks different: the card network layer is dominated by UnionPay in place of Visa, and the third-party payment layer follows a two-tier structure of merchant direct connection plus NetUnion clearing. The underlying logic is identical: a complete system of real-name KYC combined with Alipay transaction protection, WeChat Pay dispute resolution, and the 12315 complaint mechanism ensures recourse in cases of faulty goods or merchant fraud. The rail structure may differ, but the fundamental premise—that the final authority rests with a verified individual—remains the same on both sides.

The entire system operates under several assumptions that payment stack engineers have taken for granted since the 1960s:

First, people are in the loop. The fundamental assumption behind every card payment is that "the final decision is made by a person," which is why the chargeback safety net makes sense and KYC can be tied to an individual.

Second, the transaction amount must be economically viable to cover the fee. For Visa/Mastercard card rail, the standard merchant quote for an end-to-end transaction is approximately 2.9% + $0.30. This is reasonable for a $100 retail purchase, but it is not economically feasible for an API call, data endpoint, or web scraping request priced at $0.005: the fee would be 60 times higher than the transaction value itself.

Third, merchants expect buyers to be human. Their fraud detection models, inventory strategies, and customer service processes all assume the buyer is a person with an IP history, a device fingerprint, and genuine purchasing habits.

Fourth, the chargeback pipeline is properly integrated. The card rails dispute process assumes the buyer can describe "I didn't purchase this" or "it doesn't match the description." In 60 years of retail commerce, these three questions—who the buyer is, what was purchased, and what went wrong—have always been clear.

The LLM agent breaks all four assumptions simultaneously. Imagine an AI research agent completing a single task: pulling content from dozens of web pages, calling several paid data endpoints, and purchasing one or two small reports—each action costing between $0.001 and $0.10. None of these actions require a human to click "confirm," and each individual cost is far below the $0.30 card rail threshold. The merchant is not dealing with a person, and when issues arise, it’s impossible to determine "who authorized what." All four assumptions have completely failed.

AI inference services like Anthropic, OpenAI, and Replicate currently function using a "pre-top-up + internal metering" model: bundling thousands of API calls into a single large card payment to circumvent the minimum fee constraints of card rails. Once agents autonomously purchase reports, subscribe to tools, or sign API contracts on-chain, this architecture can no longer support it.

More direct signals come from content providers. According to Cloudflare’s AI Crawl Control report for August 2025: their customers’ websites collectively return over one billion HTTP 402 status codes per day [9], with publishers using this most straightforward method to tell AI bots: “Pay to access the content.”

The issue is that HTTP 402 has been reserved since the 1998 HTTP/1.1 standard but has never evolved into a widely adopted payment protocol; AI bots receive it only as a generic "access denied" error and cannot convert it into an actual payment. This new wave of protocols—x402, Pay per Crawl, and L402—aims to solve exactly this: adding a genuine payment layer on top of HTTP 402 that enables AI bots to automatically complete payments.

Within the same ecosystem, another key set of data exists: traditional search engines direct readers back to the original website when crawling pages, but AI crawlers do not. According to Cloudflare’s June 2025 statistics: Google sends back one visit for every 14 crawls, OpenAI’s crawler is at a ratio of 1,700:1, and Anthropic’s crawler reaches an astonishing 73,000:1 [9]. After scraping content, AI crawlers deliver answers directly within chat interfaces like ChatGPT and Claude, meaning users never visit the original site. The publisher’s previous revenue model—“searched → clicked → monetized via ads”—has been completely severed.

Combine these signals: The traditional payment stack cannot handle the high-frequency micropayments required by agents; AI reasoning services are forced to rely on pre-charging as a workaround; content providers receive neither payment from crawlers nor revenue from user traffic. On one side, agents must have the ability to pay; on the other, merchants/publishers must have a way to receive payment from agents—with four fundamental assumptions of traditional payment protocols standing in between. This gap is large enough that three types of players simultaneously recognized it, entered the space, and adopted three distinct approaches based on their own assessments and assets.

Three forces, three entry approaches

Three approaches correspond to three business models, each targeting a different segment of the traditional payment stack being disrupted. This section first clarifies these approaches; the engineering details of each path will be explored in depth in the next three articles.

Since the card rail cannot economically support transactions below one cent, bypass it and use stablecoins for on-chain settlement instead.

The first protocol to achieve scale on this path was x402, released by Coinbase in May 2025. It embeds USDC transfers directly into HTTP 402 responses: an agent requests a resource, the server responds with a 402 status and price, the agent signs an EIP-3009 authorization with their wallet, and Coinbase’s facilitator pays the gas fee on-chain. On the Base chain, gas costs less than $0.0001 per transaction with settlement in about 2 seconds. By April 2026, when the Linux Foundation took over, x402 had processed 165 million transactions cumulatively, generating approximately $50 million in total volume and serving 69,000 active agents, with Solana accounting for about 65% of the transaction volume [1].

This amount is modest by traditional payment industry standards. Visa processed $14.5 trillion, approximately 260 billion transactions, in its fiscal year 2025; Mastercard processed $9.2 trillion annually. The cumulative volume on x402 in its first year is roughly equivalent to what Visa processes in 2 minutes or Mastercard in 3 minutes. It may seem small, but note that x402 targets the micropayment market—segments that traditional card rails cannot serve due to minimum fee thresholds—placing it in a completely different arena from Visa and Mastercard’s retail consumer transactions.

The same concept is blossoming across other ecosystems: Lightning Labs adapted it for Bitcoin as L402 [8], Skyfire added a KYC identity layer to enable USDC microstream payments, Anthropic and other AI inference services use it for pay-per-use billing [7], Cloudflare Pay per Crawl applies the same protocol layer to charge content providers for web crawlers, with publishers like Condé Nast, TIME, and AP already integrated [9], and Circle is launching its own Agent Stack for vertical integration in May 2026.

In May 2026, AWS Bedrock AgentCore Payments launched with native support for x402, marking the first time a major public cloud provider has natively supported agent-paid transactions. Specifically for API service providers, BlockSec has integrated on-chain address labeling and Phalcon compliance risk screening as x402-paid endpoints, with a starting price of $0.10 per agent call, settled in USDC on Base [1].

The card network does not intend to let agent traffic leave its own rail, so it has chosen to adapt itself.

The representative solution is Agent Pay [3], released by Mastercard 2025-4. It does not create anything new but integrates the agent identity into its decade-old MDES tokenization infrastructure: MDES, originally used for virtual card tokens in Apple Pay and Google Pay, now adds two new fields to the token—an agent identifier and a session-scope object specifying limits, merchant scope, and expiration time. When ChatGPT or Microsoft Copilot assist users in making payments via Agent Pay, the settlement path remains the traditional card rail, with all interchange and chargeback dispute procedures fully preserved. Funds continue to flow according to the established model: card networks collect interchange and network fees, issuing banks receive the bulk of the interchange fee, and acquirers/PSPs collect processing fees. The moat lies in chargeback protection—a dispute resolution mechanism backed by decades of legal precedent—that crypto pathways still lack an equivalent for as of 2026.

Visa took a different approach: in October 2025, it partnered with Cloudflare to launch the Trusted Agent Protocol, issuing verifiable identities to agents based on HTTP Message Signature and Web Bot Auth, with 12 partners including Adyen, Checkout.com, and Worldpay joining the pilot [5]. Stripe chose a third path, transforming its existing Issuing product into a system that programmatically issues one-time virtual cards to agents [10]. Agents directly call Stripe’s Issuing API to generate virtual cards for transactions, forcibly linking the lowest fee thresholds of the card rail with agents’ high-frequency, self-funded needs.

Neither抢 rail nor issue new credentials; only standardize the format of "intent credentials" exchanged between agents and merchants.

The representative approach is AP2, released in 2025-9 by Google in collaboration with over 60 companies [2]. AP2 splits an agent’s shopping on behalf of a user into three signed steps, each producing a cryptographic credential (collectively called a Mandate) using W3C Verifiable Credentials:

1. 1. The user first signs the Intent Mandate: instructing the agent "what to do" (e.g., "find a pair of white running shoes under $200"). After signing, the user can leave, and the agent uses this authorization to search for products.
2. 2. After the agent finds the specific product, they show the shopping cart to the user; when the user confirms in real time, they sign the Cart Mandate: locking in exactly what to buy and for how much.
3. 3. At the moment the agent initiates the payment, sign the Payment Mandate: inform the downstream settlement party (card network / chain / Lightning) that "this payment was initiated by the agent on behalf of the user, and should be processed under this scope."

AP2 Three-Step Mandate Process: Step 1 — Intent Mandate (user pre-authorizes → agent uses credentials to search for products); Step 2 — Cart Mandate (agent finds product → user confirms and locks cart in real time); Step 3 — Payment Mandate (agent initiates payment → downstream Visa/x402/Lightning receives proof)

The three-layer signature forms a complete chain, allowing any downstream rail—whether Visa card, x402 USDC, or Lightning—to obtain the same cryptographic proof. AP2 itself does not settle transactions, nor does money flow through it; Google earns standard-setting power along this path: once AP2 becomes the de facto standard, all rails will build atop it. Coinbase and Lowe’s have already demonstrated a complete checkout flow using AP2 plus stablecoins [2].

Simultaneously, OpenAI and Stripe launched ACP, taking a different path: directly turning ChatGPT into a shopping interface; The Information later reported that OpenAI takes approximately a 4% platform fee from Shopify merchants [4], marking the first time an LLM provider directly collects money at the checkout layer. In March 2026, Stripe and Paradigm jointly released MPP, alongside their own Tempo chain and an IETF standardization proposal [6]; OpenAI and Anthropic were named as design partners for MPP—a first in the history of any payment protocol design.

The three pathways are not mutually exclusive. Stripe is simultaneously a founding member of the x402 Foundation, a primary advocate for ACP, a primary advocate for MPP, and a partner of AP2—its involvement in all four standards is the most direct evidence that the landscape remains unclear as of early 2026 [6].

Stripe four-way betting diagram: x402 / ACP / MPP / AP2 — Stripe places bets on all four tables simultaneously. Caption: "when you don't know who wins, just bet every table."

What’s next?

The current landscape is still unclear, meaning all three paths are still testing their own answers. Each path is backed by a complete engineering structure, business model, customer base, and regulatory position, along with its own set of security risks—these must be examined separately to make meaningful comparisons. The following articles will focus on each path individually.

First, let’s discuss the crypto-native path. How did x402 revive the 27-year-old "zombie status code," HTTP 402? What enabled Skyfire to turn KYC into identity infrastructure for the agent era and attract Anthropic? And how did Cloudflare Pay per Crawl use the same status code to build a paid crawling marketplace for publishers? Behind these three threads lies a single business insight: traditional card rails will never enter markets where each transaction is under one cent.

The third piece examines how card networks are responding. Mastercard Agent Pay embeds agent identities into its decade-old MDES tokenization infrastructure, essentially adding a new layer atop its most robust asset; Visa TAP issues agents verifiable digital IDs, aligning with the traditional intuition that “I need to know who you are before you can swipe”; Stripe Issuing for Agents provides agents with programmatically issued one-time virtual cards, forcibly connecting high-frequency microtransactions to the card rail. Behind these three approaches lies the same core question: In the agent era, can the legal definition of “cardholder” still hold up?

The fourth layer shifts to the protocol level, where the most surprising developments occur: protocols like AP2, ACP, and MPP don’t move money themselves, but are competing to establish the standard format for “intent credentials” in the agent era. AP2 chains cryptographic signatures using three Mandates, ACP turns ChatGPT directly into a checkout interface, and MPP is pushing HTTP 402 into the IETF standardization process. The winner at this layer will become the OS above all rails.

The final article provides a horizontal comparison, placing the three pathways side by side to compare identity, funds, and dispute resolution; examining how much each pathway is affected by attack scenarios unique to the agent era; exploring how the industry is filling the responsibility gap while regulators remain silent; and analyzing the seemingly contradictory phenomenon of Stripe betting on all four standards simultaneously—what industry insights does this reveal?

If you're still unfamiliar with some of these payment terms, feel free to quickly glance at the cheat sheet at the end before continuing. The remaining four articles in this series will assume you already understand what chargeback, acquirer, issuer, interchange, merchant of record, token, and Mandate mean.

Glossary

References

[1] Coinbase. "Introducing x402." May 2025. https://www.coinbase.com/developer-platform/discover/launches/x402 ; Linux Foundation. "Launching the x402 Foundation." April 2, 2026. https://www.linuxfoundation.org/press/linux-foundation-is-launching-the-x402-foundation-and-welcoming-the-contribution-of-the-x402-protocol ; BlockSec x402 paid API (address labeling + Phalcon compliance screening endpoint, starting at $0.10 per request, settled in USDC on Base). https://x402.blocksec.ai/ ; Visa Annual Report FY2025 (SEC 10-K). https://investor.visa.com/ ; Mastercard FY2025 results. https://investor.mastercard.com/

[2] Google Cloud. "Announcing Agent Payments Protocol (AP2)." September 16, 2025. https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol ; AP2 official documentation. https://ap2-protocol.org/

[3] Mastercard. "Unveils Agent Pay." April 29, 2025. https://www.mastercard.com/global/en/news-and-trends/press/2025/april/mastercard-unveils-agent-pay-pioneering-agentic-payments-technology-to-power-commerce-in-the-age-of-ai.html

[4] OpenAI. "Buy it in ChatGPT." 2025-09-29. https://openai.com/index/buy-it-in-chatgpt/ ; Agentic Commerce Protocol GitHub. https://github.com/agentic-commerce-protocol/agentic-commerce-protocol

[5] Visa. "Trusted Agent Protocol press release." October 14, 2025. https://investor.visa.com/news/news-details/2025/Visa-Introduces-Trusted-Agent-Protocol-An-Ecosystem-Led-Framework-for-AI-Commerce/default.aspx

[6] Stripe. "Developing an open standard for agentic commerce." 2026. https://stripe.com/blog/developing-an-open-standard-for-agentic-commerce ; Stripe. "Introducing the Machine Payments Protocol." 2026. https://stripe.com/blog/machine-payments-protocol ; IETF Internet-Draft "The Payment HTTP Authentication Scheme." https://datatracker.ietf.org/doc/html/draft-ryan-httpauth-payment-01 ; The Defiant. "Tempo launches mainnet, unveils Machine Payments Protocol with Stripe." 2026-03. https://thedefiant.io/news/blockchains/tempo-launches-mainnet-unveils-machine-payments-protocol-with-stripe

[7] Skyfire. https://skyfire.xyz/ ; "Skyfire Launches Open KYAPay Protocol With Agent Checkout." June 26, 2025. https://www.businesswire.com/news/home/20250626772489/en/

[8] Lightning Labs. "The Agents Are Here." February 11, 2026. https://lightning.engineering/posts/2026-02-11-ln-agent-tools/

[9] Cloudflare. "Introducing pay per crawl." 2025-07-01. https://blog.cloudflare.com/introducing-pay-per-crawl/ ; Cloudflare. "Introducing AI Crawl Control." 2025-08-28. https://blog.cloudflare.com/introducing-ai-crawl-control/ ; Cloudflare. "The crawl before the fall... of referrals." 2025-06. https://blog.cloudflare.com/ai-search-crawl-refer-ratio-on-radar/

[10] Stripe. "Giving agents the ability to pay." 2025. https://stripe.com/blog/giving-agents-the-ability-to-pay ; Stripe Issuing for agents docs. https://docs.stripe.com/issuing/agents